Project 4: Eigenfaces

Computer Vision (CSE 455), Winter 2012
Name: Daseul Lee

• Methodology
  1. Using 10 eigenfaces, find a face in elf.tga image by cropping one best face, with min_scale of 0.45, max_scale of 0.55, step of 0.1.

  main --findface elf.tga nosmile_eig10.face 0.45 0.55 0.01 crop 1 crop_elf.tga

  2. With a portrait on the web, find a face in that image by cropping one best face.

  main --findface joker.tga nosmile_eig10.face 0.5 0.7 0.05 crop 1 crop_joker.tga

  3. Find faces in "group_neutral (1).tga" by marking.

  main --findface "group\group_neutral (1).tga" nosmile_eig10.face 0.55 0.75 0.05 mark 2 mark_group_neutral1.tga

  4. With several pictures of at least 4 people, find faces in each image by marking.

  main --findface "group\group_neutral (11).tga" nosmile_eig10.face 0.55 0.75 0.05 mark 4 mark_group_neutral11.tga
  main --findface group3.tga nosmile_eig10.face 0.5 0.7 0.02 mark 20 mark_group3.tga
  main --findface "group\oldclass_neutral.tga" nosmile_eig10.face 0.65 0.85 0.05 mark 33 mark_oldclass_neutral.tga

• Sample Results
  1. Elf
     
     
  2. portrait
     
     
  3. group1
     
     
  4. group of 4 people
     
     
  5. group of random people
     
     
  6. old class picture
     
     
• Discussion
  Q1: What min_scale, max_scale, and scale step did you use for each image?

  1. Elf : 0.45, 0.55, 0.01
  2. portrait : 0.5, 0.7, 0.05
  3. group1 : 0.55, 0.7, 0.05
  4. group of 4 people : 0.55, 0.75, 0.05
  5. group of random people : 0.5, 0.7, 0.2
  6. old class picture : 0.65, 0.85, 0.05
  

  Q2: Did your attempt to find faces result in any false positives and/or false negatives? Discuss each mistake, and why you think they might have occurred.

  1. Elf : found face correctly
  2. portrait : found face correctly
  3. group1 : found face correctly
  4. group of 4 people : the hand of third person is mistaken as a face because of the contrast in regions above and below hands is similar to eyebrows of a face.
  5. group of random people : 10 out of 20 people are recognized; some people's face is not oriented toward the front or partically hidden, which might have caused the error. Some people's clothes are recognized as face because its contrast in color is similar to that in contour of a face.
  6. old class picture : 14 out of 32 people are recognized; some people's eyes are closed or their face is not oriented toward the front, causing false negatives. Also, clothes and the edge between the desk and the background is mistakenly recognized as a face for the same possible reason described above, adding to false positive.
  

• Methodology
  1. Using the cropped, non-smiling students pictures, compute 6 eigenfaces.
  

  main --eigenfaces 6 25 25 nonsmiling_cropped\list.txt nosmile_eig6.face

  2. Use the same set of images to compute a userbase.
  

  main --constructuserbase nosmile_eig6.face nonsmiling_cropped\list.txt nosmile.user

  3. Verify each user in the userbase against his/her smiling picture to find false negative, and against someone else's smiling picture to find false positive.
  

  main --verifyface smiling_cropped\smiling-1.tga nosmile.user nonsmiling_cropped/neutral-1 nosmile_eig6.face 90000.0

  main --verifyface smiling_cropped\smiling-1.tga nosmile.user nonsmiling_cropped/neutral-2 nosmile_eig6.face 90000.0

  4. Try different threshold to find the best one that results in the fewest false positives and false negatives.
  
• Sample Results
  Verifying Faces: False positive/negative using Different Threshold
  
  

  Threshold	False Negative	False Positive
  80,000	7/33	2/33
  90,000	5/33	3/33
  100,000	4/33	5/33
  110,000	4/33	7/33
  120,000	4/33	8/33

  
  
  

  Results of verifying faces using a threshold of 9000
  
  

  Original

  

  Smiling-match (false negative - Red)

  

  Smiling-nonmatch (false positive - Green)

  

  
  

  ---

  Original

  

  Smiling-match (false negative - Red)

  

  Smiling-nonmatch (false positive - Green)

  

  
  

  ---

  Original

  

  Smiling-match (false negative - Red)

  

  Smiling-nonmatch (false positive - Green)

  
• Discussion
  Q1: What MSE thresholds did you try? Which one worked best? What search method did you use to find it?

  I tried 80000, 90000, 100000, 110000 and 120000 and 90000 worked best. To find this threshold, I started out with 80000(which is a number used in a sample command in Readme.txt) and incremented the number by 10000 while calculating the number of false negatives. It decreased as the threshold increases; however, at some point the number of false negative does not change much. And starting from this threshold that did not make a big difference in false negaitve rate, I did verifyface against a picture that is not of a user to find false positive. By decrementing the threshold by 10000, I could see the false positive rate decreases but then again stablizes. After collecting data, I summed the number of false positive and false negative for each threshold and found 90000 resulted in the fewest error.

  Q2: Using the best MSE threshold, what was the false negative rate? What was the false positive rate?

  Using 90000, the false negative rate is 5/33(0.1515), and the false positive rate is 3/33(0.0909).